Identifying defect patterns in a wafer map during manufacturing is crucial to find the root cause of the underlying issue and provides valuable insights on improving yield in the foundry. Currently used methods use deep neural networks to identify the defects. These methods are generally very huge and have significant inference time. They also require GPU support to efficiently operate. All these issues make these models not fit for on-line prediction in the manufacturing foundry. In this paper, we propose an extremely simple yet effective technique to extract features from wafer images. The proposed method is extremely fast, intuitive, and non-parametric while being explainable. The experiment results show that the proposed pipeline outperforms conventional deep learning models. Our feature extraction requires no training or fine-tuning while preserving the relative shape and location of data points as revealed by our interpretability analysis.

翻译：在制造过程中识别晶圆图中的缺陷模式对于发现根本问题至关重要，并为提高晶圆厂的良率提供宝贵见解。当前使用的方法通常基于深度神经网络来识别缺陷，但这些模型普遍规模庞大且推理时间较长，还需依赖GPU支持才能高效运行。这些问题使得这些模型无法适用于晶圆制造厂的在线预测。本文提出一种极其简单但有效的晶圆图像特征提取技术。该方法极快、直观且无参数，同时具有可解释性。实验结果表明，所提出的流程优于传统深度学习模型。我们的特征提取无需训练或微调，且可解释性分析显示，该方法能够保留数据点的相对形状和位置信息。